Comparative diagnostic performance and stability of deep learning- and CFD-based CT-FFR across vessels, cardiac phases, and centers.

Although CT-derived FFR (CT-FFR) based on deep learning (DL) and computational fluid dynamics (CFD) is increasingly used for functional ischemia assessment, direct head-to-head multi-center evidence regarding their diagnostic stability in the same cohort remains limited. This study aimed to compare the diagnostic performance and robustness of DL-based versus CFD-based CT-FFR against invasive FFR across coronary branches, cardiac phases, clinical centers, and ischemia-positive gray-zone lesions. We retrospectively analyzed 220 patients (277 vessels) who underwent coronary CTA and invasive FFR from two centers. CT-FFR was calculated using representative commercial DL-based and CFD-based algorithms. Diagnostic performance was evaluated using invasive FFR as the reference standard. Subgroup analyses were performed for target vessels, reconstruction phases, imaging centers, and gray-zone lesions. DL and CFD showed high and similar diagnostic performance. The AUC was 0.90 (95% CI: 0.88-0.93) for DL and 0.89 (95% CI: 0.86-0.92) for CFD, and the difference was not significant (p > 0.05). Both methods were strongly correlated with invasive FFR (rho = 0.71 for DL; rho = 0.68 for CFD; both p < 0.001). The subgroup analyses showed stable performance across vessels, cardiac phases, and centers (all p > 0.05). In gray-zone lesions, DL and CFD showed comparable correct classification rates (86.4% vs. 84.6%, p = 0.690) and false-negative rates (13.6% vs. 15.4%, p = 0.690). DL-based and CFD-based CT-FFR showed similar and strong diagnostic performance for detecting hemodynamically significant stenosis. These findings support the potential use of both approaches as non-invasive functional assessment tools in selected patients.